FFT-Based Random Projections for Real-Time Sensor AI
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing AI/ML computing systems for controlling fault-intolerant and safety-critical apparatuses face challenges in achieving high computational efficiency and model capacity, often requiring specialized hardware and incurring high costs and environmental impacts, limiting their usability for real-time processing of sensor data.
Innovation Solution
A method involving random projections in the frequency domain using Fast Fourier Transformations (FFT) to compute output values from input values, reducing the number of operations and memory needed, enhancing computational efficiency and model capacity.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If specialized hardware (GPUs, TPUs, AI accelerators) is used for AI/ML processing, then computational efficiency and model capacity are improved, but device complexity, cost, and environmental impact increase
Solution Approach 1:
The patent replaces traditional matrix multiplication operations with frequency-domain convolution using Fast Fourier Transform (FFT). This substitution allows standard CPUs to achieve GPU-level performance for AI/ML tasks by transforming the computational problem into a domain where efficient algorithms (FFT) can be applied, thereby eliminating the need for specialized hardware while maintaining high computational efficiency
Solution Approach 2:
The patent changes the computational parameters by operating in the frequency domain rather than the spatial domain. By transforming input data and weight matrices into the frequency domain using FFT, the system leverages the efficiency of frequency-domain operations to reduce computational complexity and resource requirements, enabling high-performance AI/ML processing on general-purpose hardware
2Productivity
If the number of operations and memory requirements are reduced, then computational efficiency is improved, but model capacity may be compromised
Solution Approach 1:
The patent transitions from spatial-domain matrix operations to frequency-domain convolution operations. This dimensional transformation allows the system to achieve the same computational efficiency as highly optimized hardware architectures while maintaining full model capacity, as the FFT-based approach preserves all necessary information through the transformation process
Applied Scientific Principles
This section explains which scientific principles are used to turn an abstract innovation direction into a practical engineering solution.
Function Achieved in This Case
This approach allows for efficient AI/ML training and inference in applications previously constrained by model capacity and efficiency, enabling real-time processing and reducing resource requirements, suitable for fault-intolerant systems like vehicles and medical devices.
Implementation Method 1
defining a projection number of projections as a multiple of the input number based on a respective multiplying number; precomputing a transformed projection tensor as a transformation of a projection tensor having a projection size of the input number multiplied by the multiplying number and containing weight values in the frequency domain; obtaining a transformed embedded input tensor by computing a transformed input vector as a transformation of the input vector into the frequency domain
Data Source
Figure 1
Figure 2~5
AI summary
A method for computing at least one output value (B) from a number of input values (A) by a computing device (2), in particular for performing computer-assisted data classification and/or data regression based on real-time processing of sensor data (I), a corresponding computer program (3), computer-readable data-carrier (4), computing device (2) and apparatus (1), such as a vehicle, in particular an aircraft, are described, the method comprising the steps of providing an input number (D) representing a vector size of an input vector (x) containing the input values (A); defining a projection number (N) of projections as a multiple of the input number (D) based on a respective multiplying number (O); precomputing a transformed projection tensor (W_FD) as a transformation of a projection tensor (W) having a projection size of the input number (D) multiplied by the multiplying number (O) and containing weight values (w) in the frequency domain; obtaining a transformed embedded input tensor (X_FD) by computing a transformed input vector (x_FD) as a transformation of the input vector (x) into the frequency domain; embedding the transformed input vector (x_FD) into the transformed embedded input vector (X_FD); and using the transformed projection tensor (W_FD) and the transformed embedded input vector (X_FD) for computing the at least one output value (B).